Dissertation > Excellent graduate degree dissertation topics show

Spectral clustering method in image segmentation

Author: LiJunYing
Tutor: WangXiLi
School: Shaanxi Normal University
Course: Computer Software and Theory
Keywords: Image Segmentation Spectral clustering algorithm Semi-supervised learning Algorithm stability Large-scale image
CLC: TP391.41
Type: Master's thesis
Year: 2011
Downloads: 144
Quote: 0
Read: Download Dissertation

Abstract

The 21st century is an era of information explosion, the image has become an important means of human access to information and expression information and transmit information in people's social and economic life, as a visual image processing technology generic image processing technology, playing Yulai more important role. Image segmentation is the key to the image processing to the image analysis, is the basis of the image understanding and recognition and, therefore, the effective image segmentation means in the image processing plays a very important role. Spectral clustering algorithm is built on the basis of spectral graph theory, compared with the traditional clustering algorithm, it has the advantage of in any shape of sample space clustering and converges to the global optimal solution, and is very suitable for many practical problems , image segmentation naturally included. Therefore, in recent years, the spectral clustering algorithm for image segmentation by many scholars. Currently, image segmentation techniques based on spectral clustering algorithm has taken some good results, However, since the technology is still in the early research stage, it is there are still many of the urgent need to study and solve problems. (1) stability is not enough one of the problems faced by the spectral clustering algorithm, but also hinder the algorithm widely used in image segmentation. Spectral clustering algorithm for clustering feature vector clustering algorithm into a spectral clustering algorithm is very sensitive to the initial cluster centers, usually there will be a lack of stability shortcomings, which greatly hampered the algorithm in widely used in image segmentation. Aiming at this problem, a semi-supervised spectral clustering algorithm combined with Bayesian decision. The algorithm through distance learning methods based on Bayesian decision to adjust the contents of the similarity matrix, thus improving the distribution in order to improve the stability and accuracy of the algorithm for clustering the eigenvectors; same time, the algorithm for use with constraints K_ means clustering algorithm adjusted for clustering feature vector clustering divided, in order to further improve the stability and accuracy of the algorithm. The thinking is simple, easy to implement, while use of the structure implied by the large number of unmarked sample distribution fully the use of a small amount of labeled samples contained constraint information and category information. Experimental results show that the algorithm has a significant improvement in the stability and accuracy than the traditional spectral clustering. (2) applied to the vast amounts of data is difficult the inherent defects spectral clustering the limit spectral clustering algorithm is also widely used in the large-scale image segmentation of the important reasons. Assume that the size of the data set X for n, then the spectral clustering algorithm, the size of the similarity matrix W n2. Obviously, for large-scale problem, such as the typical image segmentation, W the amount of calculation and memory capacity are difficult to accept, and mention solving their eigenvectors. For the proposed split spectral clustering method based on the neighboring characteristics of large-scale image. In this method, through uniform sampling appropriate size of the original image to obtain a smaller image mode, also known as sample images, and then use the sample image segmentation based on of Nystrom approximation spectral clustering algorithm on, and finally the use of image segmentation results of the aforementioned sampling rules, according to some estimates, the completion of the original image, the final generic estimates. The thinking is simple, easy to implement, spectral clustering segmentation algorithm as the core, to give full play to the spectral clustering algorithm clustering in any shape of sample space and converge to the global optimum advantage of. Experiments show that the method can quickly and effectively to achieve good segmentation of large-scale complex image.

Related Dissertations

CLC: > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device
© 2012 www.DissertationTopic.Net Mobile